An AI agent is software that carries out an instruction rather than just answering it. You describe what you want, the agent plans the steps, uses the tools you make available, and delivers a result for you to review. This explanation skips the jargon, so no 'LLM orchestration' or 'tool calling.' You get what an agent does, how it works, and when it is useful for your business.
An AI agent is a program that pursues a goal through several steps. A language model does the reasoning. Around it sit the parts that let the program observe, decide, and act. Those parts are access to files and systems, memory for earlier decisions, and rules about what it may and may not do. Together, those parts turn a conversational tool into something that can carry out work.
We use AI agents in the plural because, in practice, several often run alongside each other. One agent monitors your quote workflow, another prepares the weekly report, and another keeps an eye on Search Console. Each agent has its own task, permissions, and boundaries, so you decide what is allowed for each job. Gartner predicts that 33 percent of enterprise software applications will include agentic AI by 2028, up from less than 1 percent in 2024. Major vendors use similar terms. As ServiceNow explains, an agent is a program that independently gathers data, makes decisions, and carries out tasks.
In one sentence: AI agents are digital workers that accept an instruction, work through it independently, and present the result for review. See how Voltti puts that into practice on our page about AI agents for businesses.
An AI agent works in a loop: plan, act, check, adjust. It repeats that loop until the task is done, then hands you the result to review.
First, it reads the instruction and plans the steps. Then it carries out one step at a time with the tools it has: reading a file, running a script, or querying an integration. After each step, it checks the result and decides whether to continue, adjust, or try again. Its own checks do not guarantee that everything is correct, so a person reviews the result.
Here is an example from our own work. We asked our own Hermes installation at Voltti to compile a prioritized list of articles that needed revision. Using its own read-only access, the agent read our Search Console data and calculated impressions and positions for each article. It weighed those against how current the content was and turned the list into concrete tasks. It recorded each step in a log. Search Console access is read-only; CMS access is separate. The agent edits directly in Storyblok and sometimes publishes the work itself. We then review it and make corrections where needed. One task on that list was this article.
That loop is also why an agent does more than a basic question-and-answer chatbot. The chatbot stops at the answer.
A basic chatbot gives you text and waits for your next question. An AI agent can act on your instruction: it reads, writes, runs scripts, and hands you something to review.
An agent can also retain context and earlier decisions when memory is configured. Some chat products offer tools and memory too, so the difference depends on the setup, not just the product name.
For a closer comparison, our article on AI agents versus chatbots covers the differences, costs, and a decision table. This table summarizes a basic chatbot without action tools or persistent memory versus an agent configured with those capabilities:
| Feature | Chatbot | AI agent |
|---|---|---|
Type of interaction | Text in, text out | Instruction in, result out |
File access | None in this basic setup | Yes, within agreed projects |
Memory | Within one session in this setup | Retains decisions and context when configured |
Using tools | Limited or none in this setup | Scripts, web search, integrations |
Human review | Review answers before using them | Yes, with approval for high-risk actions |
Where it works | In the chat window | Through your platforms: Telegram, Slack, desktop |
Common tasks for an agent in a business today:
Reading, summarizing, and editing files
Running scripts and checking the results
Finding information in projects and knowledge bases
Drafting emails, quotes, and replies to customers (more in our article on AI agents in customer service)
Analyzing data and preparing reports
Preparing or completing recurring workflows
The best starting point is not automating everything at once. Pick one task that comes back every week; this guide shows how to spot recurring tasks worth automating. For more on selecting suitable processes, read about automating business processes with AI.
Not every AI agent suits business work. Check five things. Is it open source, so you are not locked into one vendor? Can it run on your own infrastructure, at your business? Can it use tools, so it works with your existing software? Does it have memory, so it keeps business context? And is there support for updates, backups, and help?
We work with Hermes Agent, an open-source AI agent from Nous Research that meets these five points, with installation and support from Voltti. It works in the chat and work tools your team already uses, as our page on Hermes Agent for business shows. For a practical look at the setup, see Hermes Agent explained.
Start with one recurring task, not a strategic plan. Choose a task that takes time and causes frustration every week, and run a short pilot with one person who reviews the result. The biggest obstacle is rarely the technology; it is deciding where to start.
Agents are not a shortcut. Gartner expects over 40 percent of agentic AI projects to be canceled by the end of 2027. The reasons it gives are escalating costs, unclear business value, or inadequate risk controls. That is why we keep the first step small: one task, one reviewer, and clear permissions.
The full process is described in Voltti's pilot approach. To get ready, prepare your AI agent pilot in seven questions.
Voltti installs and supports AI agents for businesses in Belgium and the Netherlands, from the first pilot to everyday use. Wondering whether your recurring task is a good starting point? Use the button below to discuss your pilot and choose a useful first task with us.
Sharing this with Dutch-speaking colleagues or clients? There is also a Dutch version: Wat is een AI-agent? Zo werkt het, zonder jargon.
Not exactly. ChatGPT combines a chat interface with tools and agent features, depending on your plan and settings. An AI agent such as Hermes can carry out tasks with the right access: reading files, running scripts, and working within your projects. Compare the enabled capabilities and permissions, not just the names.
No. If you work on your own, you can use Hermes Desktop on your workstation, with the agent's tools in an app. A team can work with the agent through platforms it already uses, such as Telegram, Slack, or Discord. The technical setup stays behind the scenes, with support from Voltti.
The environment runs on a system at your business. Access is limited to selected employees, and the agent gets the minimum permissions needed for each task, with read-only access wherever possible. Encrypted backups go to a separate location on your own network. Read more about our approach to security and privacy.
Yes, and we recommend it. Start with one person using Hermes Desktop on a workstation and one workflow. As more people join, the team can use a channel it already works with, such as Buzz, Telegram, Slack, or Discord. There is no required channel migration. Expanding a working setup is usually easier than setting it up for the first time.
That depends on what you want to automate and where the environment runs. A simple setup without integrations costs less than one with CRM connections or custom scripts. Voltti therefore always provides a custom quote, without a standard monthly rate. To see what determines that quote, read our article on AI agent costs.
